Task 0: deterministic listing probes for both deployable harnesses

[davisbuilds]

First task of the distribution-profiles plan: the read-only probes every later task trusts. Task 0 is the plan's stop gate — nothing downstream is worth anything if the listing measurement is subtly wrong.

Adds scripts/profiles/{probe_codex,probe_claude}.py, six fixtures, and 27 hermetic tests. Suite 313 → 340. No mutation: both probes re-run on demand, so staleness is checked rather than assumed.

It works

Codex  (dojo): 41 entries, 2 namespaced, 4,132 tok = 76.0% of 5,440, absolute render mode
Claude (dojo): 94 loaded → 75 sent, demand 23,287 vs 8,000 budget, 52 descriptions removed

Four findings that change later tasks

1. Only skill lines are charged against the Codex budget. render.rs sums line_cost over entries; intro prose and section headers are not counted. Every figure this program has produced charged the whole block and overstated by ~2 points. Corrected in spec and plan: ordinary session 76% (not 78%), viral 95% (not 97%).
2. The alias roots table is a rounded difference of two whole bodies, not a sum of per-line costs — 24 tokens against 65 on the live fixture. The sum shrinks the apparent limit by 41 tokens, enough to misreport a listing that exactly fits. Task 3 must port aliased_metadata_overhead_cost, not approximate it.
3. Codex has a third degradation tier nobody had recorded: whole-skill omission, and unlike truncation it does emit a warning into the prompt. Task 3 needs three Codex shapes, not one.
4. Codex reports resolved paths. A project root symlinked into the catalog — how every dojo checkout exposes itself — appears in the roots table as the canonical path, so comparing against the unresolved cwd finds project scope nowhere and returns a confident zero. Task 4 must resolve.

Five more are in the plan's Task 0 "Executed" block, including that the plan's stated reason for reading messages was wrong: dojo's ## Available Skills decoy lives in messages too, so the section doesn't discriminate — only the literal opening sentence does. Right conclusion, wrong reason; the fixture retains both to pin it.

Vendor parity is exact, not approximate

On a capture that truncated, the port's entry cost equals limit − table_cost to the token (5,416 = 5,416), because render_lines_with_description_budget spends the description budget down to the last one. An off-by-one anywhere in the byte counting, the newline, or the ceiling division breaks it. Asserted as a test.

The mutation probe caught one that review would not have

Five deliberate breakages introduced against the detectors. Four failed immediately. One passed all 23 tests: setting sent := loaded — the filesystem-versus-context error in a new costume — because the over-budget warning line also carries a skill count and was silently overwriting sent. So sent was only ever proven on over-budget captures, which is most of them.

Fixed by keeping the two counts separate and cross-checking them, and by capturing a 1M-window session that emits no warning at all, which isolates the sent path. That fixture doubles as a direct demonstration of SC-03: identical catalog, same machine, same day — over budget on 200k, silent on 1M.

Two portability defects caught before CI

Both would have passed locally and failed every CI run — the docs/archive false-pass shape from R17 again:

• Tests referenced the capture directory by absolute path. Now rebuilt under tmp_path with a deliberately dangling symlink, since the target path is what classification compares against and it belongs to the machine that produced the capture.
• fingerprint() crashed when the harness binary was absent. Now an unknown version, which is a legitimate stale state.

Fixture hygiene

The --debug-file captures were committed raw at 252 lines, of which the parser reads 4. The remainder carried permission allow-rule lists from both settings scopes, MCP server names, and hook configuration — machine configuration with no business in a fixture, and hard to remove once in history. Trimmed to the six skill-resolution lines with a header noting what was cut (37 KB → 1.3 KB). The other four fixtures were audited and are clean; the API_KEY hits in the Codex captures are env-var names inside skill descriptions, never values.

A correction to the spec

Earlier in the day I wrote that no live foreign entry remained and the SC-05 fixture would have to be constructed — reasoning from microsoft-foundry being disabled to a general absence, without running the probe. The probe reports spreadsheet, listed in every session. Both the error and the correction are recorded in the spec and plan, since inferring an absence one command away from the instrument is this contract's own subject matter.

Verification

• pytest tests/ -q → 340 passed
• validate_spec.py PASS, validate_plan.py PASS, validate_skill_contract.py --strict PASS
• Hermetic: passes with the capture directory deleted, from a foreign cwd, and with PATH stripped of both harness binaries

🤖 Generated with Claude Code

[davisbuilds]

@codex

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9cf1539ca9

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".